Hey, thanks for the feedback. From your screenshot, it looks like multiple composer-2-fast and claude-4.6-opus-high calls were triggered while creating the plan, so the cost can climb very quickly.
Here’s what’s happening. When the agent creates a plan, it can start sub-agents like Explore and Bash, and they run in their own context. Built-in sub-agents use a faster and cheaper model by default, but the main agent still uses the model you selected. In your case, that included claude-4.6-opus-high, and it uses a lot of tokens.
Here are a couple things you can try to reduce costs:
- For planning tasks, try a lighter model like
claude-4.6-sonnetorcomposer-2.
On making sub-agent model usage more customizable, that’s a fair suggestion.
Thanks for your reply and suggestion
.
Firstly, I’m happy with the efficiency of the subagent, but from an economic perspective, I’d like to be able to control whether the cursor’s built-in subagent controls whether Composer2 is enabled fast.
Glad to hear subagent efficiency is working well for you. We’ve noted your feedback about controlling the built-in subagent model, like toggling composer-2-fast, and passed it to the team. You’re not the only one asking, it’s on our radar.
For now, the best way to manage costs is to switch to a lighter main model for planning tasks, like I mentioned. But your request for more granular subagent model control is fair, and we’re tracking it.
Please allow us to specify what model the subagents to use or simply allow us to not use subagents at all . It’s not normal to force the Composer 2 as only option, I had no issues with speed before whatsoever and this is just burning requests for the old plan 500 request based user. I had a prompt that triggered 3 subagents, so that’s 8 cents each + 4 cents for main prompt, that’s 28 cents for a single prompt. Either allow us to disable them, create a non thinking variant of composer or allow us to select a different model. I would be so fine with a slower response than to spend 7x on a prompt!
@sircoderin, got the feedback. On the legacy request-based plan, subagents in Composer-2 really do increase usage in a noticeable way. A request to let you pick the model for subagents (or disable them) is already on our radar, and we’re also tracking this specifically on the legacy request-based plan. No fix timeline yet.
For now, a couple workarounds:
- For planning and exploration, use a lighter main model like
claude-4.6-sonnetorcomposer-2, so the main call doesn’t go throughopusormax. - If you create plans often, it’s worth trying Plan Mode with a lighter model, then switch in Agent only when you’re ready to implement.
Once there’s an update on controlling subagent models, I’ll reply in the thread.